Semiconductor Process Optimization Using Filtered Neural Ensemble Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine-learning techniques for semiconductor process optimization face challenges in reliability due to output variations and sparse data spaces, leading to unpredictable results and decreased accuracy.
Innovation Solution
A machine-learning method utilizing multiple neural network models with different weights and biases, where output values are sorted and those within preset ranges are removed, and a final output value is calculated based on the remaining values, thereby reducing prediction errors and output variance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single neural network model is used for semiconductor process optimization, then the device complexity is low, but the output variation is high and reliability is poor
Solution Approach 1:
The patent combines multiple neural network models (first neural network models and second neural network model) into an integrated optimization system. The first neural network models predict multiple output values for the same input, and these are combined with a second neural network model that processes the sorted output values to generate final optimized parameters. This merging approach reduces output variation and improves prediction reliability while maintaining manageable system complexity through modular architecture.
2Reliability
If training data is split to prevent overfitting, then the neural network generalization improves, but the prediction error in sparse data space expands
Solution Approach 1:
The patent segments the training process into multiple independent neural network models trained on different randomized portions of the training dataset. Each first neural network model is trained with a different random seed, creating diverse prediction patterns. This segmentation allows the system to capture various aspects of the data distribution without requiring a single model to overfit to any particular subset, thereby maintaining accuracy in sparse data regions.
Solution Approach 2:
The patent changes the parameter initialization (random seeds) for each neural network model to create parameter diversity. By training multiple models with different random weight initializations on the same or different data subsets, the system generates a range of predictions that can be aggregated to reduce variance and improve reliability in sparse data spaces.
3Productivity
If gradient descent is applied to optimize input for low output, then the optimization direction is clear, but the output deviation according to training index increases
Solution Approach 1:
The patent implements a feedback mechanism where multiple first neural network models generate predictions, these predictions are sorted and fed into a second neural network model that produces optimized parameters. The system iteratively refines predictions by using the sorted output values from multiple models as input to the second model, which then generates optimized semiconductor parameters. This feedback loop reduces output deviation while maintaining optimization efficiency.
Data Source
AI summary
A machine-learning method for semiconductor process optimization may include inputting semiconductor-related parameters into each of first neural network models and outputting, based on the semiconductor-related parameters, a predicted figure of merit of a semiconductor device as a first output value from each of the first neural network models. After a semiconductor manufacturing process is performed with a semiconductor manufacturing parameter, electrical measurement parameter values may be measured using one or more measuring devices. The semiconductor-related parameters may include electrical measurement parameter values measured on one or more semiconductor devices. The method may also utilize a feedback loop between an output and an input of the first neural network models so that the electrical measurement parameter values can be updated based on an output value of the first neural network models. A second neural network model may also be used. A computing device and a system are also disclosed.